src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <fcntl.h>
   4 #include <sys/statvfs.h>
   5 #include <sys/vfs.h>
   6 #include <unistd.h>
   7
   8 #include "acl-util.h"
   9 #include "dirent-util.h"
  10 #include "fd-util.h"
  11 #include "fileio.h"
  12 #include "fs-util.h"
  13 #include "missing_magic.h"
  14 #include "nspawn-def.h"
  15 #include "nspawn-patch-uid.h"
  16 #include "stat-util.h"
  17 #include "stdio-util.h"
  18 #include "string-util.h"
  19 #include "strv.h"
  20 #include "user-util.h"
  21
  22 #if HAVE_ACL
  23
  24 static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
  25         acl_t acl;
  26
  27         assert(fd >= 0);
  28         assert(ret);
  29
  30         if (name) {
  31                 _cleanup_close_ int child_fd = -EBADF;
  32
  33                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  34                 if (child_fd < 0)
  35                         return -errno;
  36
  37                 acl = acl_get_file(FORMAT_PROC_FD_PATH(child_fd), type);
  38         } else if (type == ACL_TYPE_ACCESS)
  39                 acl = acl_get_fd(fd);
  40         else
  41                 acl = acl_get_file(FORMAT_PROC_FD_PATH(fd), type);
  42         if (!acl)
  43                 return -errno;
  44
  45         *ret = acl;
  46         return 0;
  47 }
  48
  49 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  50         int r;
  51
  52         assert(fd >= 0);
  53         assert(acl);
  54
  55         if (name) {
  56                 _cleanup_close_ int child_fd = -EBADF;
  57
  58                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  59                 if (child_fd < 0)
  60                         return -errno;
  61
  62                 r = acl_set_file(FORMAT_PROC_FD_PATH(child_fd), type, acl);
  63         } else if (type == ACL_TYPE_ACCESS)
  64                 r = acl_set_fd(fd, acl);
  65         else
  66                 r = acl_set_file(FORMAT_PROC_FD_PATH(fd), type, acl);
  67         if (r < 0)
  68                 return -errno;
  69
  70         return 0;
  71 }
  72
  73 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  74         _cleanup_(acl_freep) acl_t copy = NULL;
  75         acl_entry_t i;
  76         int r;
  77
  78         assert(acl);
  79         assert(ret);
  80
  81         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  82         if (r < 0)
  83                 return -errno;
  84         while (r > 0) {
  85                 uid_t *old_uid, new_uid;
  86                 bool modify = false;
  87                 acl_tag_t tag;
  88
  89                 if (acl_get_tag_type(i, &tag) < 0)
  90                         return -errno;
  91
  92                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
  93
  94                         /* We don't distinguish here between uid_t and gid_t, let's make sure the compiler checks that
  95                          * this is actually OK */
  96                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
  97
  98                         old_uid = acl_get_qualifier(i);
  99                         if (!old_uid)
 100                                 return -errno;
 101
 102                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 103                         if (!uid_is_valid(new_uid))
 104                                 return -EINVAL;
 105
 106                         modify = new_uid != *old_uid;
 107                         if (modify && !copy) {
 108                                 int n;
 109
 110                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 111                                  * beginning, so that we copy all entries, starting from the first, this time. */
 112
 113                                 n = acl_entries(acl);
 114                                 if (n < 0)
 115                                         return -errno;
 116
 117                                 copy = acl_init(n);
 118                                 if (!copy)
 119                                         return -errno;
 120
 121                                 /* Seek back to the beginning */
 122                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 123                                 if (r < 0)
 124                                         return -errno;
 125                                 continue;
 126                         }
 127                 }
 128
 129                 if (copy) {
 130                         acl_entry_t new_entry;
 131
 132                         if (acl_create_entry(&copy, &new_entry) < 0)
 133                                 return -errno;
 134
 135                         if (acl_copy_entry(new_entry, i) < 0)
 136                                 return -errno;
 137
 138                         if (modify)
 139                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 140                                         return -errno;
 141                 }
 142
 143                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 144                 if (r < 0)
 145                         return -errno;
 146         }
 147
 148         *ret = TAKE_PTR(copy);
 149
 150         return !!*ret;
 151 }
 152
 153 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 154         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 155         bool changed = false;
 156         int r;
 157
 158         assert(fd >= 0);
 159         assert(st);
 160
 161         /* ACLs are not supported on symlinks, there's no point in trying */
 162         if (S_ISLNK(st->st_mode))
 163                 return 0;
 164
 165         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 166         if (r == -EOPNOTSUPP)
 167                 return 0;
 168         if (r < 0)
 169                 return r;
 170
 171         r = shift_acl(acl, shift, &shifted);
 172         if (r < 0)
 173                 return r;
 174         if (r > 0) {
 175                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 176                 if (r < 0)
 177                         return r;
 178
 179                 changed = true;
 180         }
 181
 182         if (S_ISDIR(st->st_mode)) {
 183                 acl_free(acl);
 184
 185                 if (shifted)
 186                         acl_free(shifted);
 187
 188                 acl = shifted = NULL;
 189
 190                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 191                 if (r < 0)
 192                         return r;
 193
 194                 r = shift_acl(acl, shift, &shifted);
 195                 if (r < 0)
 196                         return r;
 197                 if (r > 0) {
 198                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 199                         if (r < 0)
 200                                 return r;
 201
 202                         changed = true;
 203                 }
 204         }
 205
 206         return changed;
 207 }
 208
 209 #else
 210
 211 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 212         return 0;
 213 }
 214
 215 #endif
 216
 217 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 218         uid_t new_uid;
 219         gid_t new_gid;
 220         bool changed = false;
 221         int r;
 222
 223         assert(fd >= 0);
 224         assert(st);
 225
 226         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 227         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 228
 229         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 230                 return -EINVAL;
 231
 232         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 233                 if (name)
 234                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 235                 else
 236                         r = fchown(fd, new_uid, new_gid);
 237                 if (r < 0)
 238                         return -errno;
 239
 240                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 241                 if (name) {
 242                         if (!S_ISLNK(st->st_mode))
 243                                 r = fchmodat(fd, name, st->st_mode, 0);
 244                         else /* Changing the mode of a symlink is not supported by Linux kernel. Don't bother. */
 245                                 r = 0;
 246                 } else
 247                         r = fchmod(fd, st->st_mode);
 248                 if (r < 0)
 249                         return -errno;
 250
 251                 changed = true;
 252         }
 253
 254         r = patch_acls(fd, name, st, shift);
 255         if (r < 0)
 256                 return r;
 257
 258         return r > 0 || changed;
 259 }
 260
 261 /*
 262  * Check if the filesystem is fully compatible with user namespaces or
 263  * UID/GID patching. Some filesystems in this list can be fully mounted inside
 264  * user namespaces, however their inodes may relate to host resources or only
 265  * valid in the global user namespace, therefore no patching should be applied.
 266  */
 267 static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
 268
 269         assert(sfs);
 270
 271         return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
 272                F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
 273                F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
 274                F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
 275                F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
 276                F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
 277                F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
 278                F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
 279                F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
 280                F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
 281                F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
 282                F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
 283                F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
 284                F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
 285                F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
 286                F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
 287 }
 288
 289 static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
 290         _cleanup_closedir_ DIR *d = NULL;
 291         bool changed = false;
 292         struct statfs sfs;
 293         int r;
 294
 295         assert(fd >= 0);
 296
 297         if (fstatfs(fd, &sfs) < 0)
 298                 return -errno;
 299
 300         /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
 301          * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
 302          * when we hit procfs, sysfs or some other special file systems. */
 303
 304         r = is_fs_fully_userns_compatible(&sfs);
 305         if (r < 0)
 306                 goto finish;
 307         if (r > 0) {
 308                 r = 0; /* don't recurse */
 309                 goto finish;
 310         }
 311
 312         /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
 313         if ((sfs.f_flags & ST_RDONLY) ||
 314             access_fd(fd, W_OK) == -EROFS)
 315                 goto read_only;
 316
 317         if (S_ISDIR(st->st_mode)) {
 318                 if (!donate_fd) {
 319                         int copy;
 320
 321                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 322                         if (copy < 0) {
 323                                 r = -errno;
 324                                 goto finish;
 325                         }
 326
 327                         fd = copy;
 328                         donate_fd = true;
 329                 }
 330
 331                 d = take_fdopendir(&fd);
 332                 if (!d) {
 333                         r = -errno;
 334                         goto finish;
 335                 }
 336
 337                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 338                         struct stat fst;
 339
 340                         if (dot_or_dot_dot(de->d_name))
 341                                 continue;
 342
 343                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 344                                 r = -errno;
 345                                 goto finish;
 346                         }
 347
 348                         if (S_ISDIR(fst.st_mode)) {
 349                                 int subdir_fd;
 350
 351                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 352                                 if (subdir_fd < 0) {
 353                                         r = -errno;
 354                                         goto finish;
 355
 356                                 }
 357
 358                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 359                                 if (r < 0)
 360                                         goto finish;
 361                                 if (r > 0)
 362                                         changed = true;
 363
 364                         } else {
 365                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 366                                 if (r < 0)
 367                                         goto finish;
 368                                 if (r > 0)
 369                                         changed = true;
 370                         }
 371                 }
 372         }
 373
 374         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 375          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 376          * tree is properly chown()ed already. */
 377         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 378         if (r == -EROFS)
 379                 goto read_only;
 380         if (r > 0)
 381                 changed = true;
 382
 383         r = changed;
 384         goto finish;
 385
 386 read_only:
 387         if (!is_toplevel) {
 388                 _cleanup_free_ char *name = NULL;
 389
 390                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 391                 (void) fd_get_path(fd, &name);
 392                 log_debug("Skipping read-only file or directory %s.", strna(name));
 393                 r = changed;
 394         }
 395
 396 finish:
 397         if (donate_fd)
 398                 safe_close(fd);
 399
 400         return r;
 401 }
 402
 403 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 404         struct stat st;
 405         int r;
 406
 407         assert(fd >= 0);
 408
 409         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 410          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 411          * following the concept that the upper 16-bit of a UID identify the container, and the lower 16-bit are the actual
 412          * UID within the container. */
 413
 414         if ((shift & 0xFFFF) != 0) {
 415                 /* We only support containers where the shift starts at a 2^16 boundary */
 416                 r = -EOPNOTSUPP;
 417                 goto finish;
 418         }
 419
 420         if (shift == UID_BUSY_BASE) {
 421                 r = -EINVAL;
 422                 goto finish;
 423         }
 424
 425         if (range != 0x10000) {
 426                 /* We only support containers with 16-bit UID ranges for the patching logic */
 427                 r = -EOPNOTSUPP;
 428                 goto finish;
 429         }
 430
 431         if (fstat(fd, &st) < 0) {
 432                 r = -errno;
 433                 goto finish;
 434         }
 435
 436         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 437                 /* We only support containers where the uid/gid container ID match */
 438                 r = -EBADE;
 439                 goto finish;
 440         }
 441
 442         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 443          * that if the top-level dir has the right upper 16-bit assigned, then everything below will have too... */
 444         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 445                 return 0;
 446
 447         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 448          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 449          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 450
 451         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 452                 if (fchown(fd,
 453                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 454                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 455                         r = -errno;
 456                         goto finish;
 457                 }
 458         }
 459
 460         return recurse_fd(fd, donate_fd, &st, shift, true);
 461
 462 finish:
 463         if (donate_fd)
 464                 safe_close(fd);
 465
 466         return r;
 467 }
 468
 469 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 470         int fd;
 471
 472         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 473         if (fd < 0)
 474                 return -errno;
 475
 476         return fd_patch_uid_internal(fd, true, shift, range);
 477 }